Re: [PATCH v3 05/15] xprtrdma: Remove last ib_reg_phys_mr() call site

[Jason Gunthorpe <jgunthorpe@obsidianresearch.com>]

On Tue, Jul 21, 2015 at 05:41:22PM -0500, Steve Wise wrote:
> On 7/20/2015 5:42 PM, Jason Gunthorpe wrote:
> >On Mon, Jul 20, 2015 at 05:41:27PM -0500, Steve Wise wrote:
> >>>B) why bother to check? Are machines with <4GB interesting, and worth
> >>>supporting a special optimization?
> >>No, but cxgb3 is still interesting to user applications, and perhaps NFSRDMA using FRMRs.
> >Doesn't look like the NFS client will work. It requires an all
> >physical memory lkey for SEND and RECV buffers..
> >
> >Jason
> 
> Looks like cxgb3 supports LOCAL_DMA_LKEY and MEM_MGT_EXTENSIONS so dma mrs
> aren't required for NFSRDMA:
> 
> t4:~/linux-2.6/drivers/infiniband/hw/cxgb3 # grep IB_DEVICE_ iwch_provider.c
>         strlcpy(dev->ibdev.name, "cxgb3_%d", IB_DEVICE_NAME_MAX);
>         dev->device_cap_flags = IB_DEVICE_LOCAL_DMA_LKEY |
>                                 IB_DEVICE_MEM_WINDOW |
>                                 IB_DEVICE_MEM_MGT_EXTENSIONS;

Neat. Is the dma_mask set properly (I don't see any set at all)?

> So cxgb3 can still support NFSRDMA and user verbs w/o get_dma_mr(). I'll
> submit a patch soon to only support get_dma_mr() if unsigned long is 4
> bytes...

So, NFS and RDS seem to be the only iWarp compatible ULPs?

NFS has worked, and will continue to work with the global lkey.

RDS looks like it relies on an insecure all physical rkey, so it won't
work until that is fixed.

So, I'd just use sizeof(physaddr_t) > 4 as the test. The only people
that could be impacted are RDS users using distro kernels on machines
with less than 4G of ram. I somehow doubt there are any of those...

Jason